DeepSeek TUI is model-agnostic at the wiring layer but shines when paired with DeepSeek V4 via API. This reference collects related families people still deploy locally or fine-tune—complete with typical sizes for discovery on Hugging Face or GGUF mirrors.
Many flagship DeepSeek checkpoints use MoE: large total parameter counts with smaller subsets active per token. That impacts VRAM planning—always read the model card for actives vs totals before downloading quantized builds.
Current-generation models DeepSeek positions for agentic coding and long context. DeepSeek TUI is tuned around V4 APIs and workflows.
| Model | Params | Context | License | Notes |
|---|---|---|---|---|
| deepseek-v4-flash | Mixture-of-experts (public docs) | 1M tokens | API terms | Cost-efficient default for tooling-heavy sessions. |
| deepseek-v4-pro | Mixture-of-experts (public docs) | 1M tokens | API terms | Stronger reasoning; higher price (discount periods may apply). |
Reasoning-focused line associated with chain-of-thought style outputs. Often compared to other reasoning models for math and logic-heavy prompts.
| Model | Params | Context | License | Notes |
|---|---|---|---|---|
| DeepSeek-R1 | 671B MoE (reported) | 128K (typical API) | MIT / variants | Flagship reasoning model; distill smaller checkpoints exist. |
| DeepSeek-R1-Distill-Qwen-1.5B | 1.5B | Varies | MIT | Ultra-small distilled line for experimentation. |
| DeepSeek-R1-Distill-Qwen-7B | 7B | Varies | MIT | Common mid-size distilled checkpoint. |
| DeepSeek-R1-Distill-Qwen-14B | 14B | Varies | MIT | Balances quality and hardware requirements. |
| DeepSeek-R1-Distill-Qwen-32B | 32B | Varies | MIT | Higher quality distilled tier. |
| DeepSeek-R1-Distill-Llama-70B | 70B | Varies | MIT | Large distilled variant for strong local runs if you have GPUs. |
Prior flagship MoE generation; still widely referenced for Hugging Face downloads and GGUF conversions.
| Model | Params | Context | License | Notes |
|---|---|---|---|---|
| DeepSeek-V3 | 671B MoE (reported) | 128K (typical) | MIT | Successor narrative continues with V4 for newest API features. |
| DeepSeek-V3-0324 | 671B MoE | 128K | MIT | Point release commonly mirrored on Hugging Face. |
Code-specialized models popular for completion and repository tasks before V-class unified lines absorbed much of that traffic.
| Model | Params | Context | License | Notes |
|---|---|---|---|---|
| deepseek-coder-1.3b-base | 1.3B | 16K | DeepSeek license | Tiny baseline for constrained hardware. |
| deepseek-coder-6.7b-base | 6.7B | 16K | DeepSeek license | Useful small coder baseline. |
| deepseek-coder-33b-base | 33B | 16K | DeepSeek license | Strong local coder if VRAM allows. |
| deepseek-coder-v2-instruct | MoE (236B total, 21B active reported) | 128K | DeepSeek license | Instruct-tuned coding MoE. |
Vision-language variants for image+text prompts; relevant when your workflow mixes screenshots, diagrams, or UI captures.
| Model | Params | Context | License | Notes |
|---|---|---|---|---|
| DeepSeek-VL2 | MoE variants | Multimodal | DeepSeek license | Check Hugging Face cards for exact variant sizes (Tiny/Small/etc.). |
| DeepSeek-VL | 7B class | Multimodal | DeepSeek license | Earlier VL line; still referenced in quantization/GGUF repos. |
GGUF, AWQ, and other quantized builds vary by maintainer. Start with the official model card, then follow Local deployment for Ollama, vLLM, or SGLang routing into DeepSeek TUI.